Reproduce “Undergraduate Physics Specialization of a 2B Language Model on a Free GPU Budget: Leak-Checked Data, Half-Strength LoRA, and Held-Out Evaluation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dasari Ranga Eswar
The authors adapt Qwen3.5-2B for undergraduate physics by training a LoRA update on short worked solutions in its non-thinking style, then applying half the update. On 1,142 held-out UGPhysics questions, accuracy rose from 15.9% to 18.9%, while truncations fell. On 210 MATH-500 questions, accuracy fell from 73.0% to 63.5%; results use one run and a rule-based judge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.